Anthropic makes its hidden Fable 5 safeguards visible

Following yesterday's thread on the hidden safeguard, Anthropic's developer account has announced a visible Opus 4.8 fallback, and The Verge reports an apology for the earlier invisible behavior. Simon Willison's read is the one to watch: the problem was never that safeguards exist, it's that developers couldn't tell when one had fired.

The Verge: Anthropic apologizes for invisible distillation guardrail

The Verge

The reporting peg for the apology: what the hidden fallback did, and why developers wanted it surfaced.

Simon Willison on the safeguards change

X / @simonw

Willison walks through what the visible fallback changes for anyone building on the model, and why predictability matters more than the safeguard itself.

A US bill to govern the data-center buildout

Axios

Rep. Bresnahan's bill is the peg: the local permitting and resource fights over AI data centers are now a federal legislative agenda.

Australia bets its economy on data-center growth

The Guardian

The Labor government pitches data centers as a growth-and-resources play, the other side of the permitting fight.

How much heat does an AI data center produce, and where are they

Al Jazeera

A map-and-numbers explainer on heat output and siting that grounds the governance argument in physical cost.

Michiel Bakker on the compute imbalance

X / @bakkermichiel

A researcher's note on how lopsided compute distribution is, and a reminder of who the buildout actually serves.

DeepMind is worried about millions of agents interacting

MIT Technology Review

Google DeepMind is funding work on what breaks when many agents interact at once, the multi-agent case for runtime controls.

A runtime-governance architecture for production agents

arXiv

Proposes a layered architecture for governing what a deployed agent is allowed to do at runtime, rather than relying on model training alone.

When agents don't comply, and who's liable

arXiv

Studies agent non-compliance and the liability questions it raises, the gap runtime authority is meant to contain.

Governing agents like employees

Forbes

Supporting color from the enterprise side: the argument for giving each agent an identity, scope, and an owner.

Arbor: autonomous research via hypothesis-tree refinement

arXiv

A framework that structures an agent's hypotheses into a tree it refines, an attempt to keep autonomous research from wandering.

SciConBench: scoring scientific-conclusion synthesis

arXiv

A benchmark and harness for whether an agent's conclusions actually follow from the evidence it gathered.

Why aggregate metrics hide long-horizon agent failures

arXiv

Argues that averaged scores mask the step-level failures that matter for long-running agents, so external control loops are needed.

Guarding against overinterpreted claims in discovery agents

arXiv

Tackles the overinterpretation problem: how a discovery agent inflates a weak result into a confident claim.

Embodied-R1.5: open-weight embodied model with a PGC framework

arXiv

Claims state-of-the-art results with open weights and datasets; treat the SOTA claim as a paper result until others reproduce it.

Asynchronous, sensor-rate control for vision-language-action models

arXiv

Drops the single synchronous clock so a robot can process and act at sensor rate, aimed at steadier physical control.

UniIntervene: cutting human labor in real-world RL

arXiv

An agentic framework that reduces how often a human has to step in during real-world reinforcement learning.

A court ruling that could reach Google's AI Overviews

Indian Express

Early legal pressure on AI-generated search answers; read it before drawing firm conclusions about the ruling's reach.